Fix CLI rendering corruption and split CLI/frontend model defaults

[akseljoonas]

Summary

• stop the CLI research overlay from running a background redraw ticker that races with streamed markdown output
• keep research sub-agent routing on direct Anthropic when the active model is direct Anthropic, and mute LiteLLM warning spam in the interactive UI
• split model defaults into explicit CLI and frontend config files, and align the frontend Claude constant with the backend's Bedrock default
• restore missing recovery helpers and patch dangling tool-result handling, doom-loop polling detection, sandbox reuse messaging, and async test support

Testing

• python -m compileall agent backend
• ./frontend/node_modules/.bin/tsc -p frontend/tsconfig.json --noEmit
• UV_CACHE_DIR=/tmp/uv-cache uv run --with pytest --with pytest-asyncio python -m pytest -q

[claude]

Claude finished @akseljoonas's task in 2m 48s —— View job

---

PR Review

• Gather context and read changed files
• Review CLI rendering fix (terminal_display.py)
• Review research tool routing changes (research_tool.py, main.py)
• Review config split (cli_agent_config.json, frontend_agent_config.json)
• Review frontend changes (ChatInput.tsx, model.ts)
• Review test coverage (test_cli_rendering.py)
• Review remaining files (session_manager.py, README.md)
• Post final review

---

No blocking issues — 2 P1
Verdict: ready to merge

What I checked

• CLI rendering fix (terminal_display.py): Removal of _ticker_task and the _tick() coroutine. Verified the display now updates only on events (start, add_call, clear). No tick means no background wake-up racing with the markdown typewriter loop. Correct.
• Research model routing (research_tool.py:217–224): Old code matched "anthropic" in main_model, routing both direct-Anthropic and Bedrock-Anthropic to bedrock/us.anthropic.claude-sonnet-4-6. New code uses startswith to split them correctly. Direct-Anthropic → anthropic/claude-sonnet-4-6; Bedrock-Anthropic → bedrock/us.anthropic.claude-sonnet-4-6. Fix is correct for the stated goal.
• Config split (cli_agent_config.json, frontend_agent_config.json): All call sites in agent/main.py and backend/session_manager.py updated. Grepped for lingering main_agent_config references — none found in Python or TypeScript.
• isClaudePath broadening (model.ts:13): Changed from startsWith('anthropic/') to includes('anthropic'). Necessary because CLAUDE_MODEL_PATH now contains bedrock/us.anthropic.claude-opus-4-6-v1. No non-Claude model in MODEL_OPTIONS contains "anthropic" in its path, so the match set is correct. The predicate is frontend-only UI (quota chip visibility); actual authz is backend _is_anthropic_model at agent.py:72 which already used "anthropic" in model_id.
• _configure_runtime_logging (main.py:58–65): Suppresses LiteLLM loggers to ERROR. Called in cli() (covers interactive path) and inside headless_main() (covers direct invocation). Double-call in headless mode is harmless since setLevel is idempotent.
• Test coverage (test_cli_rendering.py): Covers all three routing branches of _get_research_model and the no-background-task regression. Tests are correct.
• Backend AVAILABLE_MODELS (routes/agent.py:41–68): bedrock/us.anthropic.claude-opus-4-6-v1 already present — frontend constant aligns.

---

P1: configs/main_agent_config.json left behind

The old config file still exists on disk and is no longer referenced by any code path, but it has "model_name": "anthropic/claude-opus-4-6" — the pre-split direct-Anthropic value. A contributor cloning the repo who reads the old file and hand-edits it would get silent no-ops. Should be removed or at minimum a comment pointing to the two new files.

Fix this →

---

P1: Bedrock Sonnet research model ID likely missing -v1

research_tool.py:222 returns "bedrock/us.anthropic.claude-sonnet-4-6" (no -v1 suffix) for the Bedrock Anthropic research path. Every other Bedrock Anthropic model ID in the codebase carries the suffix: bedrock/us.anthropic.claude-opus-4-6-v1 in configs/frontend_agent_config.json:3, backend/routes/agent.py:50. If AWS cross-region inference requires the versioned suffix (which the Opus ID suggests), Bedrock users will get a model-not-found error from LiteLLM when a research sub-agent is spawned.

This is pre-existing (the value existed before this PR), but the PR now routes Bedrock users to this path by making it the frontend default, so it becomes load-bearing. Worth verifying against the actual Bedrock model catalog before merging.

[claude]

Claude finished @akseljoonas's task in 3m 46s —— View job

---

Claude finished @akseljoonas's task ——

---

Re-review (new commit: 0140724)

• Check if prior P1s were addressed
• Review new _patch_dangling_tool_calls rewrite (manager.py)
• Review doom_loop.py result-hash fix
• Review agent_loop.py error classification and retry schedule
• Review malformed-args recovery injection
• Review new test suite
• Post updated review

---

1 P0
Verdict: changes requested

Prior P1s (not re-flagged per policy)

Both P1s from the first review remain unaddressed: configs/main_agent_config.json is still on disk, and research_tool.py:222 still returns bedrock/us.anthropic.claude-sonnet-4-6 without the -v1 suffix.

What I checked (new commit)

• _patch_dangling_tool_calls rewrite (manager.py:253–283): Forward scan correctly inserts stubs inline at the right position (immediately before the next non-tool message) for each assistant message whose tool_calls lack a matching result. Idempotent — a second call finds the stubs already present and skips. The multi-turn case (test_orphan_in_earlier_turn_still_gets_patched) is now handled correctly.
• doom_loop.py result-hash (doom_loop.py:24,59–70): ToolCallSignature.result_hash is populated from the immediately following tool message. Polling with identical args but changing results produces signatures that differ on result_hash, so detect_identical_consecutive won't fire. The "truly stuck" case (same args + same result) still fires correctly.
• Malformed-args recovery (agent_loop.py:95–148, 269–292): _detect_repeated_malformed walks backward, skipping non-tool messages and breaking on any non-malformed tool result. The recovery prompt is injected before the next LLM call. Logic and tests look correct.
• Context overflow catch (agent_loop.py:422–423, 249–250): Both streaming and non-streaming paths now wrap non-ContextWindowExceededError context overflow exceptions before they reach the generic retry handler. Correct.
• Test coverage: All five new test files exercise the production paths they claim to cover.

---

P0: Rate-limit retry budget falls short of the stated 60 s goal; test validates a dead entry

agent_loop.py:179:

_LLM_RATE_LIMIT_RETRY_DELAYS = [15, 30, 45]  # exceed Bedrock's ~60s TPM bucket window

The retry loop is for _llm_attempt in range(_MAX_LLM_RETRIES): where _MAX_LLM_RETRIES = 3 (agent_loop.py:177). The retry condition is _llm_attempt < _MAX_LLM_RETRIES - 1 (agent_loop.py:433), i.e., _llm_attempt < 2. That means:

• Attempt 0 → wait 15 s
• Attempt 1 → wait 30 s
• Attempt 2 → condition false, raise

_LLM_RATE_LIMIT_RETRY_DELAYS[2] (45 s) is never reached. The actual retry budget is 45 s, which is below the ~60 s Bedrock TPM bucket window the comment promises to exceed.

The test at test_llm_error_classification.py:737 asserts sum(_LLM_RATE_LIMIT_RETRY_DELAYS) > 60 — this passes (90 > 60) but tests the sum of the full list, not the sum of the two entries that are actually consumed.

Minimal fix — change the schedule to two entries that together exceed 60 s:

_LLM_RATE_LIMIT_RETRY_DELAYS = [30, 60]  # exceed Bedrock's ~60s TPM bucket window

30 + 60 = 90 s. No test changes needed: sum([30, 60]) = 90 > 60 still passes, and _retry_delay_for(err, 2) is None still passes since the list has length 2.

Fix this →

[akseljoonas]

Validated the remaining review points before changing code:

• Fixed: removed the orphaned configs/main_agent_config.json.
• Fixed: shortened _LLM_RATE_LIMIT_RETRY_DELAYS to the two entries the retry loop actually consumes, while keeping the effective retry budget above 60s.
• Not applied: the Bedrock Sonnet 4.6 -v1 suggestion is incorrect. The current AWS Bedrock model card lists anthropic.claude-sonnet-4-6 with geo/global IDs like us.anthropic.claude-sonnet-4-6, so leaving bedrock/us.anthropic.claude-sonnet-4-6 unchanged is intentional.

Local verification after the update: 87 passed with UV_CACHE_DIR=/tmp/uv-cache uv run --with pytest --with pytest-asyncio python -m pytest -q.

[claude]

Claude finished @akseljoonas's task in 1m 3s —— View job

---

Re-review (new commit: 4af77ac)

LGTM
Verdict: ready to merge

Prior findings

• P0 (rate-limit retry budget) — Fixed. _LLM_RATE_LIMIT_RETRY_DELAYS = [30, 60] now has exactly _MAX_LLM_RETRIES - 1 = 2 entries, matching what the retry loop actually consumes. Budget is 90 s (> 60 s). Test at test_llm_error_classification.py:99 adds a structural guard (len(...) == _MAX_LLM_RETRIES - 1) that will break if a future diff re-introduces a dead third entry.
• P1 (orphaned config) — Fixed. configs/main_agent_config.json deleted; grepped for main_agent_config across all .py/.ts/.json/.md — zero remaining references.
• P1 (Bedrock Sonnet -v1 suffix) — Author's pushback accepted. AWS model catalog lists us.anthropic.claude-sonnet-4-6 without the -v1 suffix (unlike Opus), so research_tool.py:222 is intentionally correct.

What I checked

• agent/core/agent_loop.py:179,215–226: _retry_delay_for uses attempt_index >= len(schedule) as its guard — with a 2-entry list and a loop that only retries when _llm_attempt < 2, indices 0 and 1 are exactly the consumed entries. Logic is consistent end-to-end.
• tests/unit/test_llm_error_classification.py:99–100: Both assertions (len structural guard and sum > 60) pass. No spurious tests left over from the old 3-entry list.
• configs/: Only cli_agent_config.json and frontend_agent_config.json remain. No dangling references.